A Self-Adaptive XCS

Abstract

Self-adaptation has been used extensively to control parameters in various forms of evolutionary computation. The concept was first introduced with evolutionary strategies and it is now often used to control genetic algorithms. This paper describes the addition of a self-adaptive mutation rate and learning rate to the XCS classifier system. Self-adaptation has been used before in the strength based learning classifier system ZCS. This self-adaptive ZCS demonstrated clear performance improvements in a dynamic Woods environment and stable adaptation of its reinforcement learning parameters. In this paper experiments with XCS are carried out in Woods 2, a truncated version of the Woods 14 environment and a dynamic Woods environment. Performance of XCS in the dynamic Woods 14 environment is good with little loss of performance when the environment is perturbed. Use of an adaptive mutation rate does not help or improve on this behavior. XCS has already been shown to perform poorly in the Woods 14 environment, and other long rule chain environments. Use of an adaptive mutation rate is shown to increase performance significantly in these long rule chain environments. Attempts to also self-adapt the learning rate in Woods 14-12 fail to achieve satisfactory system performance.